An ontology that links two records without saying who could see the evidence is a story, not a graph I can defend. Today I require ACL-visible evidence for ontology relationships, and I carry ACL provenance into exported training datasets. Training data that forgets the permission context will teach the next model to ignore it.
The wrong default
The usual knowledge graph is greedy. If two documents mention the same thing, draw a line. The usual training export is greedier: strip the text, ship the slice, assume the model should see what the engineer saw. Both skip the question the overlay exists to answer. Could a principal with this grant have seen this evidence?
A court that files an exhibit the jury was never allowed to view has not built a case. I want the relationship to carry the visibility of its evidence, and I want the export to carry that same stamp.
What changed
Ontology relationships now require ACL-visible evidence. If the evidence for a link is not visible under the access rules, the relationship does not get to present itself as established. That is a governance fix, not a claim that every relationship in the store has already been re-checked. New writes must pass the rule.
Exported training datasets now carry ACL provenance with them. A slice of text without the grant that made it visible is not a training example I will stand behind. This is my export path. It is not a customer’s production training job. It is the difference between teaching a model the words and teaching it the words with the permission still attached.
I am not writing the investor outreach that landed in the same day’s commits. That is not this architecture. The architecture question is whether a graph and a training slice can still explain who was allowed to see the evidence.
ACL-visible evidence means the relationship cannot be justified by a document the principal could not have retrieved. That is slower than a greedy linker. It is the only kind of link I will later be able to explain. The export path has to carry the same discipline or the model will be trained as if permissions were decoration.
This is my ontology track and my export. I am not claiming a customer graph has been rebuilt under the new rule. New writes must pass. Old writes wait for a pass I have not run.
Training datasets leave the building. Once they leave, I cannot add the grant later. If the export does not carry ACL provenance at write time, every downstream trainer will treat my own data as if it were public. I will not train a model to forget the lock on the door.
The ontology rule is the same fact pointed inward. A link I cannot defend under a principal’s grants is a link I should not have drawn. Greedy graphs look complete. They are not reviewable.
The next test
The next test is to export a slice, read the provenance on the examples, and try to construct a relationship whose evidence is not visible — and watch it refuse. Until that is shown, this is a rule in the code, not a cleaned graph.
Historical basis: VALESKA commits 4b13bd5 and b187a48, July 31, 2026. Same-day outreach drafts are not treated as this claim.